RepoMem

RepoMem

Persistent memory for AI coding agents.
Pure Python · SQLite + FTS5 · Zero dependencies · Zero API keys · Zero telemetry · MIT

---

Your AI coding assistant forgets everything when the session ends.
RepoMem gives it a brain — one that remembers across sessions, across projects, across months.

---

Table of Contents

• What it does
• Installation
• Quick start
• Claude Code skill — /repomem
• Architecture
• MCP tools
• Observation types
• Interfaces
• Background automation
• Schema
• Configuration
• Privacy
• vs. Alternatives
• FAQ
• Contributing
• Changelog
• License
• Acknowledgements
• About

What it does

Every time a Claude Code session ends, RepoMem:

1. Extracts structured observations from the session (bugs fixed, decisions made, errors seen)
2. Stores them in a local SQLite + FTS5 database at ~/.repomem/memory.db
3. At the next session start, injects relevant context back as a Claude system message

The result: Claude arrives knowing what broke last time, what decisions were made, and what's still pending — without you having to repeat yourself.

---

Installation

git clone https://github.com/SUDARSHANCHAUDHARI/RepoMem
cd RepoMem && bash install.sh

Restart Claude Code. Done.

Also available on PyPI: pip install repomem installs the package but you still need bash install.sh to wire Claude Code hooks, MCP server, and cron jobs.

Requirements: Python 3.11+ · No other dependencies (stdlib only)

Optional semantic search — for embedding-based retrieval that matches meaning over keywords, install the extra: pip install repomem[semantic], then add --semantic to any repomem search. This is fully opt-in; the default install stays zero-dependency and FTS5 is always the fallback.

install.sh handles everything automatically:

Step	What happens
Python detection	Finds Python 3.11/3.12/3.13 automatically
Package install	Copies package to ~/.repomem/lib/
Hook wiring	Registers Stop + SessionStart hooks in Claude Code settings
MCP server	Registers MCP server in Claude Code settings
Skill install	Installs /repomem skill to ~/.claude/skills/repomem/
Cron jobs	Adds nightly reflection + weekly defrag
Health check	Verifies DB is initialized and healthy

---

Quick start

repomem status                                    # DB stats
repomem doctor                                    # health check
repomem search "null pointer crash"               # search memory
repomem search "database migration" --project my-app
repomem pending                                   # open tasks
repomem decisions                                 # architectural decisions
repomem server                                    # web UI → http://localhost:39000
repomem tui                                       # full-screen terminal UI
repomem graphify                                  # god nodes for current repo
repomem obsidian                                  # export to Obsidian vault
repomem obsidian --no-wikilinks                   # export without wikilinks
repomem import-chat ~/Downloads/session.md        # import old chat export
repomem import-chat ~/Downloads/session.md --move # import + delete source
repomem resolve-error 3                           # mark error #3 resolved
repomem merge-branch feat/my-feature --pr-number 42
repomem sync --export                             # export for cross-machine sync

---

Claude Code skill — /repomem

RepoMem ships a native Claude Code skill at skills/repomem/. Once installed, type /repomem in any Claude Code session to get a full memory interface without touching the CLI.

Install the skill:

install.sh handles this automatically. If you need to reinstall manually:

mkdir -p ~/.claude/skills/repomem
cp skills/repomem/SKILL.md ~/.claude/skills/repomem/

Triggers:

You type	What happens
/repomem	Status + recent observations for current project
"save this to memory"	Saves current context as an observation
"what did we fix before?"	Searches memory for past bugfixes
"add pending task"	Adds a task to the pending queue
"show decisions"	Lists architectural decisions
"open web UI"	Starts repomem server → http://localhost:39000
"sync memory"	Runs repomem sync --export

The skill is Claude Code-only — it uses Claude Code's SKILL.md format. The underlying RepoMem tool is AI-agnostic and works with any agent that can run CLI commands.

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Claude Code session                      │
└────────────────┬───────────────────────────┬────────────────┘
                 │ session ends               │ session starts
                 ▼                            ▼
        ┌──────────────┐             ┌──────────────────┐
        │  Stop hook   │             │ SessionStart hook │
        │ capture.py   │             │   inject.py       │
        └──────┬───────┘             └────────┬─────────┘
               │ write                         │ read + rank
               ▼                               ▼
        ┌──────────────────────────────────────────────┐
        │           ~/.repomem/memory.db               │
        │  SQLite + FTS5 · 11 tables · local only      │
        └──────┬───────────────────────────────────────┘
               │
    ┌──────────┼──────────────┐
    ▼          ▼              ▼
 reflect.py  defrag.py    MCP server
 (2am daily) (Sun 3am)  (mid-session)

What happens at session end:

Step	Detail
Extract observations	Regex patterns detect bugfix/decision/upgrade/warning/learning/error/pending
Tag topics	Keywords match build, networking, database, compose, etc. — short keywords use word-boundary matching to prevent false positives
Link entities	PascalCase classes, known libraries, filenames extracted and linked
Detect errors	Exception/Crash/FAILED patterns captured with root cause + fix
Detect releases	released v1.2.0, uploaded AAB, TestFlight build, git tag v, versionName, bumped to patterns
Track branch	git branch --show-current saved per project
Store	All written to SQLite with FTS5 triggers for instant search

What gets injected at session start (priority order):

Priority	Content	Cap
1	Global + project architectural decisions	up to 20
2	Open pending tasks for this project	up to 5
3	Recent observations ranked by recency × confidence	up to 10
4	Conflict warnings (contradicting decisions)	up to 4
5	Release age warning (if >90 days since last release)	1 line
6	Cross-project patterns (seen in 2+ projects)	up to 3
7	Graphify god nodes (high-blast-radius files)	up to 5
8	Hot entities (most frequently touched classes/files, ≥3 mentions)	up to 3
9	Unresolved errors with fixes	up to 3
—	Hard cap: ≤2000 characters total	—

---

MCP tools

When an MCP client connects to RepoMem's server, 8 tools become available mid-session:

Tool	Arguments	What it does
repomem_search	query, project?, type?, limit?	FTS5 full-text search across all observations
repomem_answer	question, project?, limit?	Grounded, #id-cited memory block to answer from — no LLM call
repomem_save	type, summary, detail?, topic?, project?	Persist an observation immediately
repomem_context	project?	Full project context (decisions + pending + recent obs)
repomem_pending	project?	List all open tasks
repomem_decisions	scope?	Architectural decisions (global or project-scoped)
repomem_add_pending	task, project?, priority?	Add a pending task to the queue
repomem_resolve	id	Mark a pending task as resolved

---

Observation types

RepoMem captures 8 structured types automatically from session text:

Type	Trigger phrases	What gets stored
bugfix	"fixed crash", "resolved issue", "repaired bug"	What was broken + how it was fixed
decision	"decided to use", "switched to", "migrated to"	The choice made and why
upgrade	"upgraded to", "bumped version", "updated"	Library/tool and new version
warning	"watch out", "never", "avoid", "caution"	The thing to remember not to do
learning	"learned", "discovered", "found out", "note:"	New knowledge gained
pending	"todo", "next:", "still need", "needs to"	Work left for next session
pattern	Auto-promoted from obs seen in 3+ projects	Reusable solution across codebases
error	Exception / Crash / FAILED / fatal	Error text + root cause + fix

---

Interfaces

CLI — 22 commands

Command	Description
repomem search <q>	FTS5 full-text search, optional --project, --type, --limit, --semantic
repomem answer <q>	Grounded, #id-cited memory block for a question (no LLM call)
repomem mcp-config -c <client>	Print MCP config snippet for claude/cursor/windsurf/cline/codex
repomem add	Manually add an observation (--type, --summary, --detail)
repomem pending	List open tasks, optional --project
repomem add-pending <task>	Add a task with --priority P1/P2/P3
repomem resolve <id>	Mark a pending task resolved
repomem resolve-error <id>	Mark a tracked error as resolved
repomem decisions	List architectural decisions
repomem add-decision	Record a decision with --topic, --scope, --reason
repomem status	DB stats: obs count, sessions, projects, DB size
repomem doctor	Health: FTS5 sync, conflicts, unresolved errors, orphans
repomem entities	List extracted entities; --name to find obs by entity
repomem releases	List auto-detected releases
repomem branches	List tracked open branches
repomem merge-branch <branch>	Mark a branch as merged (--pr-number, --pr-url)
repomem import-chat <file.md>	Import a raw Claude session export into memory (--project, --move)
repomem sync	--export to git chunk, --import to merge peers
repomem obsidian	Export to Obsidian vault (--project, --vault, --no-wikilinks)
repomem graphify	God nodes for current repo, --threshold edges
repomem tui	Full-screen terminal UI
repomem server	Start web viewer, default port 39000

Web viewer

repomem server    # → http://localhost:39000

Page	Content
Dashboard	Stats overview + recent observations + unresolved errors
Observations	Browsable list with type badge, project, date
Decisions	All decisions by scope (global / project)
Pending	Task queue sorted by priority (P1 → P3)
Errors	Unresolved errors, recurrence count, known fix
Projects	All projects, observation count, last activity date
Search	Full-text search across all observations

Dark mode · No build step · Press / to focus search from any page

Terminal UI

repomem tui

Key	Action
j / ↓	Move cursor down
k / ↑	Move cursor up
Enter	Open detail view
/	Enter search mode
Esc	Exit search
o	Switch to Observations
p	Switch to Pending
d	Switch to Decisions
e	Switch to Errors
r	Refresh data
q	Quit

---

Background automation

Job	Schedule	What it does
reflect.py	2am daily	Dedup near-identical obs (80% similarity), promote cross-project patterns (3+ projects), flag contradictions, apply confidence decay (90d, <0.5 confidence)
defrag.py	Sunday 3am	Merge dupes (85% similarity), archive stale low-confidence obs, trim oversized detail to 2000 chars, rebuild FTS5 index, vacuum DB

---

Schema

11 tables across 4 phases of development (plus 1 FTS5 virtual table with 4 internal shadow tables managed by SQLite):

Table	Phase	Key columns
sessions	1	id, project, folder, repo_path, started_at, ended_at, obs_count
observations	1	type, topic, summary, detail, confidence, seen_count, is_stale, conflict_id
observations_fts	1	FTS5 virtual table (auto-synced via INSERT/UPDATE/DELETE triggers)
decisions	1	scope (ALL / project), topic, decision, reason, is_superseded
pending	1	project, task, priority (P1/P2/P3), resolved_at
patterns	1	topic, title, solution, seen_in (projects), seen_count
entities	2	name, type (class/file/library/god_node), mention_count
entity_links	2	entity_id ↔ obs_id many-to-many join
errors	2	error_text, root_cause, fix, recurred, is_resolved
releases	3	project, version_name, version_code, released_at, store
branches	3	project, branch, status (open/merged), pr_number, merged_at

See SCHEMA.md for full column-level documentation.

---

Configuration

Environment variable	Default	Purpose
REPOMEM_DIR	~/.repomem	Override storage location
REPOMEM_INSTALL	~/.repomem/lib	Override package path (used by hooks)
REPOMEM_PROJECT	auto (git remote)	Override project name detection
REPOMEM_OBSIDIAN_VAULT	~/.repomem/exports/obsidian	Override Obsidian vault export path
REPOMEM_PORT	39000	Web viewer port

---

Privacy

Guarantee	Detail
Local only	All data in ~/.repomem/memory.db — never leaves your machine
No network	Zero outbound connections, ever
No telemetry	No usage data, no crash reports, nothing
No API keys	No OpenAI, Anthropic, or any external service needed
Private tags	<private>…</private> content stripped before storage
Relocatable	REPOMEM_DIR=/your/path moves the entire database

---

vs. Alternatives

Short summary — see COMPARISON.md for 5 detailed comparison tables:

	claude-mem	Engram	Basic Memory	mem0	RepoMem
Zero dependencies	❌ npm+Bun	❌ Go binary	❌ uv+Python	❌ pip+LLM API	✅ stdlib only
Zero API keys	✅	✅	✅	❌ needs LLM	✅
One-command install	✅ npx	✅ brew	✅ uv	❌	✅ bash
Stop hook (auto-capture)	✅	✅	❌	❌	✅
SessionStart auto-injection	✅	✅	❌	❌	✅
Structured typed observations	Partial	✅	✅ categorical	❌	✅ 8 types
Per-project filtering	✅	✅	✅	❌	✅
Sleep-time reflection	❌	❌	❌	❌	✅ nightly
Conflict detection	❌	Beta (needs LLM)	❌	❌	✅ autonomous
Error tracking	❌	❌	❌	❌	✅
MCP server	✅ 4 tools	✅ 19 tools	✅	❌	✅ 8 tools
Web viewer (local, no signup)	✅ local	❌ cloud only	❌ cloud only	❌	✅ local
Terminal UI	❌	✅	Partial	❌	✅
Obsidian sync	❌	Beta	✅	❌	✅
Git cross-machine sync	❌	✅	Manual	❌	✅
DB schema migrations	❌	❌ confirmed	✅ Alembic	❌	✅
Private content tagging	✅ <private>	❌	❌	❌	✅ <private>
Fully auditable (no binary)	❌ compiled JS	❌ Go binary	✅	✅	✅
License	Apache 2	MIT	AGPL	Apache 2	MIT

memanto — the closest philosophical opposite

memanto (Moorcheh AI, paper) is a research-backed competitor worth contrasting directly, because it makes the opposite design bet at almost every turn.

	memanto	RepoMem
Retrieval	Semantic — proprietary Moorcheh engine	Lexical — SQLite FTS5
Zero dependencies	❌ Docker + Ollama (local) or cloud	✅ stdlib only
Zero API keys	Local only (cloud needs key)	✅ always
Memory model	Queryable — "not injectable" by design	Auto-injected at SessionStart
answer (grounded QA)	✅ LLM-generated	✅ agent-grounded, no LLM call
Typed observations	✅ 13 categories	✅ 8 dev-focused types
Conflict detection	✅	✅
Document ingestion	✅ pdf/docx/xlsx/csv	Markdown session import only
Research + benchmarks	✅ LongMemEval 89.8%, LoCoMo 87.1%	❌
Fully auditable	Partial — Moorcheh engine is proprietary	✅ read every line
License	MIT	MIT

The honest tradeoff. memanto wins on retrieval quality: its information-theoretic engine matches "the auth thing" to "OAuth token refresh" where FTS5 needs shared keywords. It also has a paper and benchmark numbers RepoMem can't claim. RepoMem wins on footprint and trust: nothing to run — no Docker daemon, no Ollama model pull, no engine you can't read. Just python3 and a SQLite file, fully air-gapped. memanto is built to be asked; RepoMem is built to show up automatically at session start. Pick the bet that fits how you work.

Cells above are drawn from memanto's README and paper, not from running it — a few of its capabilities (Stop-hook capture, per-project scoping, MCP specifics) were left out rather than guessed.

---

FAQ

Does it work with agents other than Claude Code?
The core tool is AI-agnostic — any agent that can run bash commands can use the CLI. The MCP server works with any MCP-compatible agent (Cursor, Gemini CLI, etc.). The /repomem skill and automatic Stop/SessionStart hooks are Claude Code-specific. Without Claude Code, use the CLI to add observations manually. See INSTALL.md.

Does it slow down session start?
No. The SessionStart hook has a 10-second timeout and injects ≤2000 characters. In practice it runs in under 1 second on a warm DB.

What happens if the DB is corrupted or missing?
All hooks exit 0 and fail silently — a missing or broken DB never blocks a Claude Code session. Run repomem doctor to diagnose and bash install.sh to reinitialize.

How do I stop it capturing something sensitive?
Wrap it in <private>…</private> in your session text. RepoMem strips that content before writing to the DB. Alternatively, set REPOMEM_PROJECT=skip to disable capture for a session entirely.

Can I use it across multiple machines?
Yes. repomem sync --export writes a JSON chunk to ~/.repomem/sync/ and commits via git. On another machine, repomem sync --import merges peer chunks. A watermark ensures only new observations are exported each run.

How do I update?

cd ~/path/to/RepoMem && git pull && bash install.sh

install.sh is idempotent — skips already-wired hooks and MCP entries, and runs DB schema migrations automatically.

---

Contributing

Contributions are welcome! Please read CONTRIBUTING.md before opening a PR.

• Report a bug
• Request a feature
• Ask a question
• Security vulnerabilities — please report privately

---

Changelog

See CHANGELOG.md for the full history.

---

License

MIT — see LICENSE for details.

---

Acknowledgements

RepoMem was built by studying the work of others. Full credit where it's due:

Project	Author / Org	What RepoMem borrowed
claude-mem	thedotmack	<private> content tagging, progressive disclosure injection strategy, Stop + SessionStart hook wiring pattern
Engram	Gentleman-Programming	SQLite + FTS5 as the right storage backend for developer memory, MCP server design, session lifecycle (start/end) tooling
Basic Memory	basicmachines-co	Sleep-time reflection concept (nightly background processing), per-project scoping, archive-never-delete philosophy
basic-memory-skills	basicmachines-co	Memory skill design patterns, wikilink-based cross-referencing, schema validation approach
agentmemory	jayzeng	Injection priority ordering, hard 2000-char context cap, scratchpad-as-highest-priority pattern, graceful degradation on hook failure
Letta / MemGPT	letta-ai	Tiered memory model (hot/cold), sleep-time compute concept, intentional write pattern
Cognee	topoteretes	Knowledge graph integration concept — led directly to the graphify integration in Phase 4
Zep	getzep	Temporal reasoning: age-labelled observations ("3mo ago"), recency × confidence ranking
mem0	mem0ai	ADD-only accumulation model (never overwrite), entity linking design, temporal decay concept
claude-code-memory-setup	lucasrosati	Vault-aware wikilink strategy (scan real notes, code-block-safe, first-occurrence-only, longest-match-first), SHORT_KEYWORDS word-boundary fix for topic detection, import-chat pipeline concept, dynamic Obsidian frontmatter tags, processed: timestamp field

See COMPARISON.md for a full breakdown of all tools studied.

---

About

I'm Sudarshan Chaudhari, a Senior Quality Engineer, Test Automation specialist, and AI systems builder based in Bangkok, Thailand.

I have 13+ years of experience in software quality engineering, working across SaaS, fintech, gaming, web, mobile, cloud, and digital signage platforms. My background combines hands-on test automation with QA leadership, test strategy, CI/CD, release quality, production investigation, and cross-platform validation.

Alongside my professional QA career, I run SudarshanTechLabs, my independent engineering and product lab where I design, build, test, and ship software across Android, web, AI, cybersecurity, developer tooling, and cross-platform applications.

What I work on

• ⚙️ Quality Engineering & Test Automation — Playwright, Selenium, Cypress, Appium, API testing, automation frameworks, end-to-end testing, CI/CD, release gates, GitHub Actions, risk-based testing, and production validation
• 🤖 AI Systems & Automation — AI agents, multi-agent orchestration, MCP servers, AI-assisted QA, prompt tooling, developer workflows, automation systems, and Claude Code plugins
• 📱 Mobile & Cross-Platform Applications — Android applications built with Kotlin and Jetpack Compose, Google Play releases, automated build and publishing pipelines, and cross-platform development spanning iOS, web, Windows, and macOS
• 🌐 Web Applications & Platforms — Full-stack applications using Next.js, TypeScript, Firebase, Cloudflare, REST APIs, and modern web infrastructure
• 🛠️ Developer Tooling & CLI Engineering — Rust, Python, TypeScript, CLI utilities, multi-repository tooling, build automation, release tooling, and engineering productivity systems
• 🛡️ Cybersecurity & Observability — Threat detection, log analysis, security auditing, vulnerability assessment, monitoring, and security-focused developer tools
• 📺 Digital Signage & Device Platforms — Content validation, playback testing, device compatibility, production investigation, monitoring, and QA across diverse hardware and operating-system environments

My work sits at the intersection of quality engineering, automation, AI, and software development. I approach products with a QA mindset from the beginning: understanding failure modes, designing for testability, automating repetitive work, and building release confidence into the engineering process.

Through SudarshanTechLabs, I also build products and tools from idea to production, covering architecture, development, testing, CI/CD, release automation, monitoring, and ongoing maintenance.

🌐 sudarshantechlabs.com · 💼 LinkedIn · 🐙 GitHub · ✉️ sunny.sudarshan@gmail.com